Fix fp32_ln for various models (#41605)

* Add is_causal to KosmosTextAttention

* Move get target_dtype to be imported elsewhere

* Fix fp32 flash attention bug in bark

* Fix is_causal in mllama

* Fix fp32 issue on StableLM

* Fix repo-consistency

src/transformers/integrations/flash_attention.py

@@ -11,6 +11,19 @@ logger = logging.get_logger(__name__)
 _use_top_left_mask = flash_attn_supports_top_left_mask()
 
 
+def get_target_dtype(query: torch.Tensor, module: torch.nn.Module) -> torch.dtype:
+    """If the query is in float32, return a target dtype compatible with flash attention. Return None otherwise."""
+    if query.dtype == torch.float32:
+        if torch.is_autocast_enabled():
+            return torch.get_autocast_gpu_dtype()
+        # Handle the case where the model is quantized
+        elif hasattr(module.config, "_pre_quantization_dtype"):
+            return module.config._pre_quantization_dtype
+        else:
+            return next(layer for layer in module.modules() if isinstance(layer, torch.nn.Linear)).weight.dtype
+    return None
+
+
 def flash_attention_forward(
     module: torch.nn.Module,
     query: torch.Tensor,
@@ -48,15 +61,7 @@ def flash_attention_forward(
     # cast them back in the correct dtype just to be sure everything works as expected.
     # This might slowdown training & inference so it is recommended to not cast the LayerNorms
     # in fp32. (usually our RMSNorm modules handle it correctly)
-    target_dtype = None
-    if query.dtype == torch.float32:
-        if torch.is_autocast_enabled():
-            target_dtype = torch.get_autocast_gpu_dtype()
-        # Handle the case where the model is quantized
-        elif hasattr(module.config, "_pre_quantization_dtype"):
-            target_dtype = module.config._pre_quantization_dtype
-        else:
-            target_dtype = next(layer for layer in module.modules() if isinstance(layer, torch.nn.Linear)).weight.dtype
+    target_dtype = get_target_dtype(query, module)
 
     # Instead of relying on the value set in the module directly, we use the is_causal passed in kwargs if it is presented
     is_causal = kwargs.pop("is_causal", None)

src/transformers/models/bark/modeling_bark.py

@@ -57,6 +57,7 @@ from .generation_configuration_bark import (
 
 
 if is_flash_attn_available():
+    from ...integrations.flash_attention import get_target_dtype
     from ...modeling_flash_attention_utils import _flash_attention_forward
 
 
@@ -78,6 +79,7 @@ class BarkSelfAttention(nn.Module):
         self.embed_dim = config.hidden_size
         self.num_heads = config.num_heads
         self.head_dim = self.embed_dim // self.num_heads
+        self.config = config
 
         if config.hidden_size % config.num_heads != 0:
             raise ValueError(
@@ -228,6 +230,8 @@ class BarkSelfFlashAttention2(BarkSelfAttention):
         if past_key_values is not None:
             key, value = past_key_values.update(key, value, self.layer_idx, {"cache_position": cache_position})
 
+        target_dtype = get_target_dtype(query, self)  # if the query is in float32, this is the dtype to cast to for FA
+
         attn_output = _flash_attention_forward(
             query,
             key,
@@ -237,6 +241,7 @@ class BarkSelfFlashAttention2(BarkSelfAttention):
             dropout=self.dropout if self.training else 0.0,
             use_top_left_mask=self._flash_attn_uses_top_left_mask,
             is_causal=self.is_causal,
+            target_dtype=target_dtype,
         )
 
         attn_output = self._merge_heads(attn_output, self.num_heads, self.head_dim)

src/transformers/models/blt/modeling_blt.py

@@ -280,12 +280,12 @@ class BltSelfAttention(nn.Module):
         self.scaling = self.head_dim**-0.5
         self.rope_theta = config.rope_theta
         self.layer_idx = layer_idx
+        self.is_causal = True
 
         self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
         self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
         self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)
         self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
-        self.is_causal = True
 
     def forward(
         self,

src/transformers/models/kosmos2/modeling_kosmos2.py

@@ -680,6 +680,7 @@ class KosmosTextAttention(nn.Module):
         self.num_heads = num_heads
         self.dropout = dropout
         self.head_dim = embed_dim // num_heads
+        self.is_causal = True
 
         if (self.head_dim * num_heads) != self.embed_dim:
             raise ValueError(

src/transformers/models/mllama/modeling_mllama.py

@@ -519,6 +519,7 @@ class MllamaTextSelfAttention(nn.Module):
         self.scaling = self.head_dim**-0.5
         self.rope_theta = config.rope_theta
         self.layer_idx = layer_idx
+        self.is_causal = True
 
         self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=False)
         self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=False)

src/transformers/models/stablelm/modeling_stablelm.py

@@ -52,6 +52,7 @@ if is_torch_flex_attn_available():
 
 
 if is_flash_attn_available():
+    from ...integrations.flash_attention import get_target_dtype
     from ...modeling_flash_attention_utils import _flash_attention_forward
 
 
@@ -495,6 +496,8 @@ class StableLmFlashAttention2(StableLmAttention):
 
         dropout_rate = self.attention_dropout.p if self.training else 0.0
 
+        target_dtype = get_target_dtype(query_states, self)
+
         attn_output = _flash_attention_forward(
             query_states,
             key_states,
@@ -505,6 +508,7 @@ class StableLmFlashAttention2(StableLmAttention):
             dropout=dropout_rate,
             use_top_left_mask=self._flash_attn_uses_top_left_mask,
             is_causal=self.is_causal,
+            target_dtype=target_dtype,
         )
 
         attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()